← Back to Voice models
Liquid AI
LFM3-Audio-2B
Released Oct 27, 2026
Liquid AI's small open-weight audio model. It takes speech or text and streams back text, speech or both, aimed at on-device voice chat.
- Price per minute
- —
- Pricing
- Open weights; Liquid AI API console linked but no price found.
- Inputs
- audio, text
- Tool calling
- Yes
- Languages
- 18
- Voices
- —
- Big Bench Audio
- —
- Access
- Generally available
- Licence
- LFM Open License v1.0
Strengths & weaknesses
Strengths
- Time to first token as low as 50 ms (causal encoder prefills while you speak)
- About 2B parameters, in BF16, GGUF, MLX and ONNX
- Turns spoken requests into tool calls
Weaknesses
- Voice chat is English only, though it understands 18 languages
- Not on the AA leaderboard
- No inference provider hosts it
Not notable: not on AA. Included as a speech-in, speech-out streaming model, but the model card doesn't call it full duplex; drop it if the directory requires full-duplex realtime. languages = 18 (the header figure; the list has 19 entries because it separates British and American English). The licence record matches the LFM Open License v1.0 entry already in models.json.
Links
API model strings
- huggingface
LiquidAI/LFM3-Audio-2B